Anatolian Journal of Computer Sciences Volume:5 No:1 2020 © Anatolian Science pp:1-13 ISSN:2548-1304 Heart Sound Classification for Murmur Abnormality Detection Using an Ensemble Approach Based on Traditional Classifiers and Feature Sets Ali Fatih Gündüz1*, Ali Karcı2 *1Akçadağ Voc. Highschool, Malatya Turgut Ozal Uni., Malatya, Turkey ([email protected]) 2Department of Computer Eng., Inonu University, Malatya, Turkey, ([email protected]) Received Date: Jan. 7, 2020. Acceptance Date: Mar. 21, 2020. Published Date : Jun. 1, 2020. Abstract— Phonocardiography (PCG) is a method based on examination of mechanical sounds coming from heart during its regular contraction/relaxation activities such as opening and closing of the valves and blood turbulence towards vessels and heart chambers. Today there are high technology tools to record those sounds in electronic environment and enable us to analyze them in detail. The constraints such as human’s limited audible range, environment noise and inexperience of physicians can be overcome by the use of those tools and development of state-of-art signal processing and machine learning methods. In this study we examined heart sounds and classified them as normal or abnormal focusing on the efficiency of ensemble classifiers. Features of heart sounds are extracted by using Discrete Wavelet Transform (DWT), Mel-Frequency Cepstral Coefficients (MFCC) and time-domain morphological characteristics of the signals. As a novel contribution, Karcı entropy is derived from DWT of PCG signals and used for the first time as feature in heart sound classification. K-Nearest Neighbor (kNN), Support Vector Machine (SVM), Multilayer Perceptron (MLP) classifiers and their ensembles are used as classifiers. Then the ensemble classifiers’ predictions based on distinct feature vectors are combined and an ensemble classifier built from team of ensemble classifiers. Classification performances of singular classifiers, single level ensemble classifiers and final ensemble classifier are compared and better results are obtained by the proposed method. Keywords : Karcı entropy, signal processing, data mining, classification, heart sound 1.Introduction and Related Work Heart and blood vessels related disorders are called as cardiovascular diseases (CVDs) which include arrhythmia, heart attacks and strokes. According to World Health Organization (WHO), CVDs are the most common causes of death globally and most people die due to CVDs annually than any other causes [1]. Even though exact causes of CVDs are not conceded clearly by authorities, there are some accepted risk factors such as hypertension, diabetes, obesity, smoking and family history of having CVD. People having the risk factors need to be more careful and concerned about their well-being. Early identification of heart diseases with high accuracy provides vital information and can be extremely important for survival of those people [2]. Doppler-Echocardiography, Magnetic Resonance Imaging (MRI), Electrocardiography (ECG) and Phonocardiogram (PCG) are effective tools for measuring dysfunctions of heart. Those methods provide accurate results although they require expensive equipment which can be used by expert physicians. On 1 the other hand, auscultation is a classical method which is defined as listening to the cardiac sounds with a stethoscope. Auscultation is simple, practical, noninvasive and cost-effective therefore it is more suitable for home and primary health care. A lot of cardiac diseases are identified firstly by finding presences of abnormalities in the heart sounds [3] however, differentiating pathological symptoms from normal heart sounds requires knowledge and experience. Moreover, it can be a subjective task even among clinicians [4, 5]. In order to obtain more successful and objective evaluation of heart sounds, physicians make use of computer-based automatic systems for heart sound diagnosis in clinics. Thanks to development of technology, today there are portable heart sound evaluation devices corroborated with digital signal analysis methods. Many state-of-art methods [6-10] have been proposed to meet this need. Those tools can be used to educate and train cardiology students and inexperienced physicians. Also, they can be very helpful in rural areas and in homecare where it is not easy to reach expert consultation [2]. During cardiac cycle, a healthy heart generates a regular pair of sounds called S1 and S2 at each pulsation due to rhythmic opening and closing of heart valves [11]. S1 and S2 are also known as fundamental heart sounds and are described as lub-dup in layman’s terms. Variation in their durations and intensities are regarded as important implications of cardiac disorders [8]. The third heart sound S3, also known as ventricular gallop, follows S2 and in most adults it is soft and subdued. However, it can be loud enough to be heard among children, pregnant females and well-trained athletes without indicating any cardiac abnormality. On the other hand, another heart sound S4, which occurs just before S1, must be inaudible in healthy subjects. Moreover, in some pathological cases those sounds are accompanied by a noise which called as heart murmurs. Having heart murmurs is not a disease but obviously indicates an abnormal turbulence of blood flow either inside heart or in vessels. Sometimes, its presence may be normal without pointing out any cardiac pathology in which case it is called as “innocent murmur” [3,12]. Detecting and classifying murmurs during auscultation is an important task for a physician. Murmurs can be caused either from inborn structural defects of the heart or from physiologic conditions which take place subsequently. Structural defects related with heart valves include stenosis (heart valve flaps restrict the blood flow by fusing together), regurgitation (flaps do not close properly allowing back flow of blood) and atresia (valves do not develop congenitally preventing it from opening) [13]. On the other hand, pregnancy, fever, anemia or post-operation effects can be counted as external causes of heart murmurs occurring later. There are various types of murmurs and they can take place in different locations of a pulse depending on their kind. Sometimes being sub-audible notwithstanding, they can be detected and seen in PCG recordings [13]. As a gold standard for grading heart murmur intensity, murmurs are graded on a 6-point level scale (Levine scale) by clinicians [14]. Separating heart sounds from murmurs can be challenging since they involve low frequency components which are hardly audible [15]. Especially presence of environmental noise, breathing sounds, rustling of the microphone due to movement make it harder for naked human ear. Therefore, there is a need for developing human-machine interfaces to help cardiologists understand and interpret PCG outputs more accurately and easily during the diagnosis [16]. Actually, automated classification of heart sounds has been studied for decades however, achieving more accurate classification results is still attracting researchers [10]. Munia et. al. [17] used Support Vector Machine (SVM) for classifying heart sounds into normal and abnormal classes by comparing five different kernel functions. Koçyiğit [18] proposed a classification method based on obtaining sub-bands of heart sounds by using DWT. Features are extracted from those sub-bands and then dimensions of feature vectors are reduced by using principal component analysis on them. Those features are given to Naïve Bayes and SVM classifiers to obtain fourteen different heart sound classes. In another study, Uzunhisarcıklı [19] examined the blood flow velocity variation due to mitral valve stenosis by performing an analysis on Doppler ultrasound signals and extracted nonlinear features from the heart sound for this purpose. Potes et. al. [7] proposed an ensemble classifier method combined from AdaBoost and Convolutional Neural Network (CNN) which are trained with time-domain and frequency-domain features. Zabihi et. al. [8] used an ensemble of 20 feedforward Neural Networks which are built from two hidden layers 2 containing 25 neurons. They used Linear Predictive Coefficients (LPC), entropy, Mel-Frequency Cepstral Coefficients (MFCC), Discrete Wavelet Transform (DWT) and Power Spectral Density (PSD) to extract features. In another study, Kay and Agarwal [9] proposed a Neural Network based method using Continuous Wavelet Transform (CWT) and MFCC for feature extraction. In feature vectors, they also used time-domain features obtained from the segmented signal such as mean, standard deviation and inter-segment ratios. Each regular pulse of heart generates a cardiac cycle which is created from distinct temporal regions. The time interval from completion of S1 to initialization of S2 is called as systole and S2 to S1 period is known as diastole [20]. Partitioning the signal into repeating individual segments is an important step of automated heart sound analysis which can be achieved by using threshold-based change point detection methods or probabilistic models [21]. Moreover, fundamental heart sound segmentation can be done by using extra reference signals such as ECG when it is recorded in parallel to PCG. Indicating cardiac cycles by R peaks of QRS complexes, ECG makes it easy to describe temporal segments of the PCG record. Segmentation can also be achieved by using different envelope extraction methods like Hilbert-Huang transform [22], Shannon energy [6] and autocorrelation [23]. Without